fix(work-driver): four grouping rules are dead on real issues — --json single-line bodies, wrong R3 literal, R4 vs title prefixes, R2 path shape

[randomm]

Context

Four of the grouping rules in work-driver-grouping.ts are documented, tested against synthetic fixtures, and do not work on real issues. Verified by running the production regexes over the live bodies of #279, #287, #288, #289, #290, #366 and #368.

What is broken

1. R3 and R4 can never fire on the multi-issue path.

commands.ts:250 fetches bodies with gh issue view <n> --json title,body,labels and stores the raw stdout verbatim (:262), which is handed to groupIssues (:265). That stdout is one line of compact JSON — newlines are two-character \n escapes, not real newlines.

Both rules are ^-anchored under /m:

• R3 work-driver-grouping.ts:126 — /^(?:Split\s*:\s*(?:true|yes|separate)|This work must ship separately)/im
• R4 work-driver-grouping.ts:205 — /^(?:title:\s*)?\[([a-z0-9_-]+)\]\s/i against firstLine

Under a single-line JSON blob the only line-start is index 0, which is {"body":". Measured: R3split=false on all 7 sampled bodies. The failure is silent — notes simply never mentions R3 or R4.

Note the explore step does this correctly: work-driver-explore.ts:72 uses plain gh issue view <n> and gets real newlines.

2. The R3 literal does not match what is actually written.

Four of four sampled operator issues contain, at line start: This work must ship as its own separate PR, independent of any other open issue. The regex literal is This work must ship separately. Even with #1 fixed, R3 stays dead for every issue in this backlog.

3. R4 is structurally impossible for /plan-authored issues.

R4 requires [tag] as the first characters of the title. /plan mandates feat: / Bug: / EPIC: / chore: / research: prefixes (pi-prompts/plan.md:36-40). The two conventions are mutually exclusive. The subsystem is in fact already encoded — as the conventional-commit scope, e.g. fix(work-driver): ….

4. R2 extracts zero paths from most real issues.

work-driver-grouping.ts:170-171 requires a directory component before the extension. Issues reference bare module names with line anchors — work-driver.ts:1274, commands.ts:262, worktree.ts:19-27.

Measured path extraction: #287 → 0, #288 → 0, #289 → 0, #290 → 0, #366 → 0, #368 → 1, #279 → 5. So groups[id].paths comes back empty for 5 of 7, and that empty scope propagates downstream.

5. docs/troubleshooting.md:680 is stale and would mislead anyone writing an issue by hand: it documents split, separate PRs, independent — the pre-#312 behaviour. grouping.ts:120-124 records that bare "independent" was removed as a false-positive magnet.

Acceptance criteria

• groupIssues receives the parsed .body, not gh issue view --json stdout; ^-anchored rules see real newlines
• R3 matches the phrasing actually used (This work must ship as its own separate PR) as well as the existing Split: true|yes|separate form
• R4 accepts a subsystem tag carried as a conventional-commit scope (fix(work-driver): → work-driver) in addition to the leading [tag] form
• R2 accepts bare module.ts:NNN anchors alongside directory-qualified paths, without matching prose
• A regression test runs all four rules against the real bodies of Verify-full tier (.pi/verify-cmd-full at ci step) + deterministic type-widening scan routed to lens context #279/feat(work-driver): repoRoot is never a dev tree — always-worktree branch mechanization + integration at repoRoot only #287/feat(concurrency): global spawn semaphore, multi-cycle work-widget, cycle attribution, jittered lens backoff #288/feat(work): parallel group queue with bounded pool + repoRoot integration mutex #289/feat(plan): decomposition-for-parallelism doctrine + deterministic plan-quality re-dispatch #290/fix(work-driver): every 429 is treated as fatal — the provider's requested retry delay is never parsed (69% of observed failures) #366/fix(work-driver): one failed group halts the entire queue — park it and continue instead #368 committed as fixtures, and asserts R3/R4 fire where they should and path extraction is non-empty
• docs/troubleshooting.md grouping section matches the implemented rules
• bun run build + cd extension && bunx tsc --noEmit && bun run check + all offline smoke tests pass

Technical context

• extension/src/commands.ts:245-265 — the --json fetch and bodiesByIssue assembly
• extension/src/work-driver-grouping.ts — R1 :154, R2 :170, R3 :126, R4 :205
• extension/src/work-driver-explore.ts:72 — the correct plain-gh fetch, for reference

Out of scope

• Intent resolution / normalised spec — separate issue; this one only repairs the fast path
• Any change to /plan or to how issues are written
• GitHub's native issue-dependency API as an R1 replacement — worth doing, but a separate concern

This work must ship as its own separate PR, independent of any other open issue.